Study finds pretraining loss can predict RL scaling before saturation
heghbalz · x · 2026-07-21
Pretraining loss appears to predict RL scaling in a non-saturated regime
The thread reports a quantitative finding: in the non-saturated regime, pretraining loss predicts post-RL pass@1 at fixed RL compute.
- The authors fit a local log-linear law.
- Reward slope increases roughly linearly with log pretraining tokens.
- The attached plots show strong correlations between pretraining metrics and observed RL reward slopes.
Related event: New Research Proposes Joint Scaling Law for Pretraining and RL(18 posts)→
More from Research
- MIT's Ziming Liu: AutoResearch doesn't generalize — blind training tricks cause 'forking' disasters — ZimingLiu11 · 2026-10-26
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26